Skip to main content

Improvements to Conversation Summary endpoint

Feature RequestsAPI

The conversation summary is marginally useful the way it’s working.  The biggest problem is that it doesn’t get the previous summary to work with, so when the conversation gets longer than its context window it becomes less than helpful as an add-on to the existing conversation.

My suggestions are:

1) The input should end with the last summary block.  This way the AI can add onto that block and incorporate features from the beginning of the conversation as well.

2) The input should include the System Prompt because then it will know what it doesn’t have to put in the summary (any character detail that originated in the System Prompt can be left out.

3) The conversation summary should concentrate on the big “beats” of the conversation.  It should look for big changes in the characters, emotional moments, promises given and fulfilled, identifying any new characters that were introduced.

4) The conversation summary should give the character creators a way to force certain things to be reflected in the summary.  For example, if the character has defined “memory” saving commands (like “[MEMORY] This character is an ass.”) then the summary should keep any memory items that it can (possibly removing ones that are duplicates or were overridden).  I tried using a construct like “(OOC: Summarizer, please include any [MEMORY] text in the summary as these are defined as data needing to be remembered.)” and it worked to some extent, but it could have been incorporated in the summary better.

1 comment

Log in to comment and vote

Comments1

  • Tan Peach

    •

    Jan 8

    This is the prompt I use to summaries my conversations:

    I have this extended conversation from VeniceAI "Character". Here's what I need. I need a detailed summary of all that has occurred, summarized in a way that will allow me to use it as the first prompt in a brand-new conversation with the same character, bringing it up to speed instantly. The goals are these:

    1. The AI gets an idea of everything that has happened in the few weeks of game time.

    2. The AI has highlights of important conversations, especially those where specific plans for the future were made (even if they aren't terribly important, they are touch stones) and especially if there are callbacks to them from later in the narrative.

    3. The results must be formatted as a complete, self-contained instruction packet for the character AI. This means the summary itself must be preceded by the exact preamble of meta-commands required to signal to the AI that this is a context ingestion, not a conversational prompt. The preamble should instruct the AI to process the text as memory, avoid out-of-character commentary, and wait for the next prompt.

    4. Note that there may be [META], [DEBUG], and [MEMORY] commands in the narrative that may be used as indicators of what is important. The [MEMORY] entries found in the story should be copied (after deduplication) to the output, since these are, by definition, important and they should be output in the order they were found in the narrative. When generating the memory section *each memory should begin with [MEMORY] for maximum impact.

    5. Include a "Final Narrative Pickup Point" section at the end of the summary. This section must contain the last 2-3 conversational exchanges (user prompt and AI response) verbatim. This serves as the exact script to be used for creating a seamless narrative bridge when the summary is ingested by the character AI, ensuring no time skip or loss of conversational context.

    6. IMPORTANT: If there is a summary at the top from even older conversations, then add your new summary to that without losing any fidelity from the original summary.

    Here is the Description:

    Here are the complete System Prompt/Instructions:

    The conversation to summarize is attached. Generate the updated summary enclosed in a single five backtick markdown code block for easy copy/pasting.